Scaling language models with more data, compute and parameters has driven significant progress in natural language processing. For example, thanks to scaling, GPT-3 was able to achieve strong results on in-context learning tasks. However, training these large dense models requires significant amounts of computing resources. In this paper, we propose and develop a family of language models named \glam (\textbf{G}eneralist \textbf{La}nguage \textbf{M}odel), which uses a sparsely activated mixture-of-experts architecture to scale the model capacity while also incurring substantially less training cost compared to dense variants. The largest \glam has 1.2 trillion parameters, which is approximately 7x larger than GPT-3. It consumes only 1/3 of the energy used to train GPT-3 and requires half of the computation flops for inference, while still achieving better overall fewshot performance across 29 NLP tasks.

Due to the computational cost of running inference for a neural network, the need to deploy the inferential steps on a third party's compute environment or hardware is common. If the third party is not fully trusted, it is desirable to obfuscate the nature of the inputs and outputs, so that the third party can not easily determine what specific task is being performed. Provably secure protocols for leveraging an untrusted party exist but are too computational demanding to run in practice. We instead explore a different strategy of fast, heuristic security that we call \textit{Connectionist Symbolic Pseudo Secrets}. By leveraging Holographic Reduced Representations (HRRs), we create a neural network with a pseudo-encryption style defense that empirically shows robustness to attack, even under threat models that unrealistically favor the adversary.

This paper proposes a self-supervised objective for learning representations that localize objects under occlusion - a property known as object permanence. A central question is the choice of learning signal in cases of total occlusion. Rather than directly supervising the locations of invisible objects, we propose a self-supervised objective that requires neither human annotation, nor assumptions about object dynamics. We show that object permanence can emerge by optimizing for temporal coherence of memory: we fit a Markov walk along a space-time graph of memories, where the states in each time step are non-Markovian features from a sequence encoder. This leads to a memory representation that stores occluded objects and predicts their motion, to better localize them. The resulting model outperforms existing approaches on several datasets of increasing complexity and realism, despite requiring minimal supervision, and hence being broadly applicable.

Exploiting similar and sharper scene patches in spatio-temporal neighborhoods is critical for video deblurring. However, CNN-based methods show limitations in capturing long-range dependencies and modeling non-local self-similarity. In this paper, we propose a novel framework, Flow-Guided Sparse Transformer (FGST), for video deblurring. In FGST, we customize a self-attention module, Flow-Guided Sparse Window-based Multi-head Self-Attention (FGSW-MSA). For each $query$ element on the blurry reference frame, FGSW-MSA enjoys the guidance of the estimated optical flow to globally sample spatially sparse yet highly related $key$ elements corresponding to the same scene patch in neighboring frames. Besides, we present a Recurrent Embedding (RE) mechanism to transfer information from past frames and strengthen long-range temporal dependencies. Comprehensive experiments demonstrate that our proposed FGST outperforms state-of-the-art (SOTA) methods on both DVD and GOPRO datasets and yields visually pleasant results in real video deblurring. https://github.com/linjing7/VR-Baseline

We present a robust learning algorithm to detect and handle collisions in 3D deforming meshes. We first train a neural network to detect collisions and then use a numerical optimization algorithm to resolve penetrations guided by the network. Our learned collision handler can resolve collisions for unseen, high-dimensional meshes with thousands of vertices. To obtain stable network performance in such large and unseen spaces, we apply active learning by progressively inserting new collision data based on the network inferences. We automatically label these new data using an analytical collision detector and progressively fine-tune our detection networks. We evaluate our method for collision handling of complex, 3D meshes coming from several datasets with different shapes and topologies, including datasets corresponding to dressed and undressed human poses, cloth simulations, and human hand poses acquired using multi-view capture systems.

The current standard approach to scaling transformer language models trains each model size from a different random initialization. As an alternative, we consider a staged training setup that begins with a small model and incrementally increases the amount of compute used for training by applying a "growth operator" to increase the model depth and width. By initializing each stage with the output of the previous one, the training process effectively re-uses the compute from prior stages and becomes more efficient. Our growth operators each take as input the entire training state (including model parameters, optimizer state, learning rate schedule, etc.) and output a new training state from which training continues. We identify two important properties of these growth operators, namely that they preserve both the loss and the ``training dynamics'' after applying the operator. While the loss-preserving property has been discussed previously, to the best of our knowledge this work is the first to identify the importance of preserving the training dynamics (the rate of decrease of the loss during training). To find the optimal schedule for stages, we use the scaling laws from (Kaplan et al., 2020) to find a precise schedule that gives the most compute saving by starting a new stage when training efficiency starts decreasing. We empirically validate our growth operators and staged training for autoregressive language models, showing up to 22% compute savings compared to a strong baseline trained from scratch. Our code is available at https://github.com/allenai/staged-training.

This work is concerned with the following fundamental question in scientific machine learning: Can deep-learning-based methods solve noise-free inverse problems to near-perfect accuracy? Positive evidence is provided for the first time, focusing on a prototypical computed tomography (CT) setup. We demonstrate that an iterative end-to-end network scheme enables reconstructions close to numerical precision, comparable to classical compressed sensing strategies. Our results build on our winning submission to the recent AAPM DL-Sparse-View CT Challenge. Its goal was to identify the state-of-the-art in solving the sparse-view CT inverse problem with data-driven techniques. A specific difficulty of the challenge setup was that the precise forward model remained unknown to the participants. Therefore, a key feature of our approach was to initially estimate the unknown fanbeam geometry in a data-driven calibration step. Apart from an in-depth analysis of our methodology, we also demonstrate its state-of-the-art performance on the open-access real-world dataset LoDoPaB CT.

We present a simple and effective self-supervised learning approach for speech recognition. The approach learns a model to predict the masked speech signals, in the form of discrete labels generated with a random-projection quantizer. In particular the quantizer projects speech inputs with a randomly initialized matrix, and does a nearest-neighbor lookup in a randomly-initialized codebook. Neither the matrix nor the codebook are updated during self-supervised learning. Since the random-projection quantizer is not trained and is separated from the speech recognition model, the design makes the approach flexible and is compatible with universal speech recognition architecture. On LibriSpeech our approach achieves similar word-error-rates as previous work using self-supervised learning with non-streaming models, and provides lower word-error-rates than previous work with streaming models. On multilingual tasks the approach also provides significant improvement over wav2vec 2.0 and w2v-BERT.

Multiscale feature hierarchies have been witnessed the success in the computer vision area. This further motivates researchers to design multiscale Transformer for natural language processing, mostly based on the self-attention mechanism. For example, restricting the receptive field across heads or extracting local fine-grained features via convolutions. However, most of existing works directly modeled local features but ignored the word-boundary information. This results in redundant and ambiguous attention distributions, which lacks of interpretability. In this work, we define those scales in different linguistic units, including sub-words, words and phrases. We built a multiscale Transformer model by establishing relationships among scales based on word-boundary information and phrase-level prior knowledge. The proposed \textbf{U}niversal \textbf{M}ulti\textbf{S}cale \textbf{T}ransformer, namely \textsc{Umst}, was evaluated on two sequence generation tasks. Notably, it yielded consistent performance gains over the strong baseline on several test sets without sacrificing the efficiency.

Semi-supervised learning (SSL) has been widely explored in recent years, and it is an effective way of leveraging unlabeled data to reduce the reliance on labeled data. In this work, we adjust neural processes (NPs) to the semi-supervised image classification task, resulting in a new method named NP-Match. NP-Match is suited to this task for two reasons. Firstly, NP-Match implicitly compares data points when making predictions, and as a result, the prediction of each unlabeled data point is affected by the labeled data points that are similar to it, which improves the quality of pseudolabels. Secondly, NP-Match is able to estimate uncertainty that can be used as a tool for selecting unlabeled samples with reliable pseudo-labels. Compared with uncertainty-based SSL methods implemented with Monte Carlo (MC) dropout, NP-Match estimates uncertainty with much less computational overhead, which can save time at both the training and the testing phases. We conducted extensive experiments on four public datasets, and NP-Match outperforms state-of-theart (SOTA) results or achieves competitive results on them, which shows the effectiveness of NPMatch and its potential for SSL.

Random Reshuffling (RR), also known as Stochastic Gradient Descent (SGD) without replacement, is a popular and theoretically grounded method for finite-sum minimization. We propose two new algorithms: Proximal and Federated Random Reshuffling (ProxRR and FedRR). The first algorithm, ProxRR, solves composite finite-sum minimization problems in which the objective is the sum of a (potentially non-smooth) convex regularizer and an average of $n$ smooth objectives. ProxRR evaluates the proximal operator once per epoch only. When the proximal operator is expensive to compute, this small difference makes ProxRR up to $n$ times faster than algorithms that evaluate the proximal operator in every iteration, such as proximal (stochastic) gradient descent. We give examples of practical optimization tasks where the proximal operator is difficult to compute and ProxRR has a clear advantage. One such task is federated or distributed optimization, where the evaluation of the proximal operator corresponds to communication across the network. We obtain our second algorithm, FedRR, as a special case of ProxRR applied to federated optimization, and prove it has a smaller communication footprint than either distributed gradient descent or Local SGD. Our theory covers both constant and decreasing stepsizes, and allows for importance resampling schemes that can improve conditioning, which may be of independent interest. Our theory covers both convex and nonconvex regimes. Finally, we corroborate our results with experiments on real data sets.

We consider two federated learning algorithms for training partially personalized models, where the shared and personal parameters are updated either simultaneously or alternately on the devices. Both algorithms have been proposed in the literature, but their convergence properties are not fully understood, especially for the alternating variant. We provide convergence analyses of both algorithms in the general nonconvex setting with partial participation and delineate the regime where one dominates the other. Our experiments on real-world image, text, and speech datasets demonstrate that (a) partial personalization can obtain most of the benefits of full model personalization with a small fraction of personal parameters, and, (b) the alternating update algorithm outperforms the simultaneous update algorithm by a small but consistent margin.

In modern machine learning systems, distributed algorithms are deployed across applications to ensure data privacy and optimal utilization of computational resources. This work offers a fresh perspective to model, analyze, and design distributed optimization algorithms through the lens of stochastic multi-rate feedback control. We show that a substantial class of distributed algorithms---including popular Gradient Tracking for decentralized learning, and FedPD and Scaffold for federated learning---can be modeled as a certain discrete-time stochastic feedback-control system, possibly with multiple sampling rates. This key observation allows us to develop a generic framework to analyze the convergence of the entire algorithm class. It also enables one to easily add desirable features such as differential privacy guarantees, or to deal with practical settings such as partial agent participation, communication compression, and imperfect communication in algorithm design and analysis.

We develop a general framework unifying several gradient-based stochastic optimization methods for empirical risk minimization problems both in centralized and distributed scenarios. The framework hinges on the introduction of an augmented graph consisting of nodes modeling the samples and edges modeling both the inter-device communication and intra-device stochastic gradient computation. By designing properly the topology of the augmented graph, we are able to recover as special cases the renowned Local-SGD and DSGD algorithms, and provide a unified perspective for variance-reduction (VR) and gradient-tracking (GT) methods such as SAGA, Local-SVRG and GT-SAGA. We also provide a unified convergence analysis for smooth and (strongly) convex objectives relying on a proper structured Lyapunov function, and the obtained rate can recover the best known results for many existing algorithms. The rate results further reveal that VR and GT methods can effectively eliminate data heterogeneity within and across devices, respectively, enabling the exact convergence of the algorithm to the optimal solution. Numerical experiments confirm the findings in this paper.

This work is concerned with the overdetermined linear least-squares problem for large scale data. We generalize the iterative Hessian sketching (IHS) algorithm and propose a new sketching framework named iterative double sketching (IDS) which uses approximations for both the gradient and the Hessian in each iteration. To understand the behavior of the IDS algorithm and choose the optimal hyperparameters, we derive the exact limit of the conditional prediction error of the IDS algorithm in the setting of Gaussian sketching. Guided by this theoretical result, we propose an efficient IDS algorithm via a new class of sequentially related sketching matrices. We give a non-asymptotic analysis of this efficient IDS algorithm which shows that the proposed algorithm achieves the state-of-the-art trade-off between accuracy and efficiency.